Facial expressions serve as a crucial medium for human emotional communication. Based on duration and intensity, expressions can be categorized into macro-expressions and micro-expressions. Characterized by their involuntary and fleeting nature (lasting less than 0.5 seconds), micro-expressions (MEs) hold significant value in fields such as business negotiations, criminal investigations, and clinical diagnosis due to their uncontrollable properties.
Traditional micro-expression recognition methods primarily rely on onset-apex frames or fixed-length sequences, often limited by insufficient utilization of temporal information. A study by Professor Haifeng Li's team at Harbin Institute of Technology, published in Frontiers of Computer Science, introduces an innovative solution modeling the dynamic evolution of micro-expressions to significantly improve recognition performance. This approach preserves richer temporal variation features through complete sequence analysis, marking a technological breakthrough in the field. The code is available at https://github.com/hitheyuhong/TA_MER_FOR_CCAC.git.
The research team pioneered a self-attention based micro-expression temporal feature analysis network:
- Dynamic Evolution Modeling: A five-layer Transformer architecture dynamically allocates attention based on the contribution of facial local movements to emotion recognition, effectively capturing long-range temporal dependencies.
- Noise Suppression Technology: Combines optical flow fields to detect subtle muscle movements, incorporating facial correction techniques and AU-based ROI localization to improve feature signal-to-noise ratio.
- Performance Superiority: Achieves state-of-the-art results on mainstream databases including CAS(ME)3 and DFME. Notably, it set a new record in the DFME seven-classification task (currently the largest ME database) with an F1-score of 0.40, securing the first place in the Automatic Micro-Expression Recognition competition of the 4th Chinese Conference on Affective Computing.
The team announced plans to focus on developing unsupervised learning methods for ME feature extractors. By leveraging massive unlabeled data, this direction aims to enhance model noise immunity and address current overfitting challenges caused by limited annotated samples, potentially benefiting high-robustness scenarios like clinical diagnosis and security monitoring.
DOI:10.1007/s11704-025-40976-3